Part of the AI in finance series.
AI literacy in group finance is not about writing better prompts. It is the ability to interrogate a model's output the way you would interrogate a capable junior's first draft: knowing what data the model was given, knowing what it could not have known, and applying proportionate professional scepticism before anything is signed or boarded.
Why does literacy decide whether the faster use cases actually pay back?
A Gartner survey of 160 senior finance leaders, run January to April 2026 and reported by CPA Practice Advisor on 24 September 2026, found that low AI literacy is now the most significant barrier finance leaders must address. The same research identified that data extraction, accounts payable and receivable automation, and report creation generally deliver returns within nine to ten months, while data management, insight generation and forecasting take longer. Those faster use cases only pay back if the output is acted on correctly. Unreviewed output that reaches a board pack or an intercompany ledger is either rework or risk - and rework in consolidation is rarely cheap.
What does a literate reviewer actually check?
A literate reviewer runs four checks before treating AI output as complete.
- Data provenance. What source data was the model given? Was it trial-balance data or a downstream summary? Were intercompany balances included or stripped? Was the cut-off correct? The model cannot flag what it was not given.
- Group-specific blind spots. A model has no knowledge of your intercompany agreements, your top-side adjustments, the entity history behind an acquired subsidiary, or the judgement that shaped last year's numbers. If the prior-year figure carries a one-off reclass, the model will trend it as normal. The reviewer must supply that context.
- Plausibility against the known shape of the group. Does the output make sense given which entities are in or out of scope this period? Does a margin movement align with what the controller for that segment reported? AI output that is directionally wrong but superficially formatted is the highest-risk category.
- What the model was asked to assume. Assumptions embedded in a prompt - a fixed FX rate, a particular elimination method, a revenue recognition policy - will not always be visible in the output. The reviewer must know what instructions were given and whether those instructions were appropriate.
How does the skill differ by role?
The reviewer's responsibility varies by where in the process a person sits, and capability-building should reflect that.
| Role | Primary risk to review for | What literacy looks like in practice |
|---|---|---|
| Group controller | Elimination errors, scope exceptions, top-side adjustments not reflected | Can reconstruct what data set the model saw; checks output against the manual elimination schedule |
| Consolidation team | Intercompany mismatches, minority interest miscalculation, FX translation errors | Runs a plausibility check against the prior period; escalates any variance the model did not flag |
| FP&A | Trend extrapolation that ignores known structural changes; forecast assumptions baked silently into a prompt | Documents what the model was and was not given before sharing output; flags where human judgement overrode the output and why |
| CFO signing | Board-ready output that carries embedded errors from earlier in the chain | Asks the preparer two questions: what did you give the model, and what did you change? Treats a blank answer as a control failure |
How do you build this capability in the team?
Three practices build reviewer literacy faster than training courses.
Review drills against known answers. Take a closed period - one where you know the correct consolidated output - and run it through your AI tooling. Ask the team to identify every deviation, every missing adjustment, every assumption the model silently embedded. Doing this with a result you can verify builds the habit of structured scepticism without the pressure of a live close.
Pairing on first-draft review. Pair a senior reviewer with a junior preparer during the first three to five periods of any new AI-assisted process. The senior reviewer narrates what they are checking and why. This transfers the mental checklist faster than documentation alone, and it surfaces group-specific context that often lives only in one person's head.
Writing down what the model was and was not given. Make it a team norm - and eventually a control - to record in a brief note alongside any AI output: the data source, the date range, any assumptions passed in the prompt, and any known limitations. This note is the reviewer's working paper. It also makes the review itself faster, because the next person in the chain does not have to reconstruct the inputs from scratch.
How do you tell the capability is working?
The leading indicator is the quality of the reviewer's challenge, not the absence of errors. A team that is building genuine literacy will start asking specific, answerable questions about model inputs before they query the output. They will write down their reasoning when they override or adjust AI output. They will escalate faster when something does not reconcile, because they understand why it might not. The lagging indicator is simpler: the ratio of AI-assisted outputs that require material correction after the first review should fall over three to six periods. If it is not falling, the review process is not working - and that is a literacy problem, not a tooling problem.
For the governance structures that sit around AI-assisted processes - including audit trails and incentive alignment - see AI governance: audit trails in finance and AI governance: incentives in the finance function. For the specific question of drafting versus signing AI-assisted board commentary, see AI variance commentary for the board pack. If your close process is not yet ready to support AI-assisted consolidation, why AI finance close pilots stall covers the readiness conditions.
Common questions
What does AI literacy mean for a group finance team?
AI literacy in group finance is the reviewer's skill: knowing what data a model was given, understanding what it could not know about the group - such as intercompany agreements, top-side adjustments, and entity history - and applying structured scrutiny to its output before it is used. It is distinct from prompting skill. A Gartner survey of 160 senior finance leaders, reported by CPA Practice Advisor on 24 September 2026, identified low AI literacy as the most significant barrier finance leaders must now address.
Which finance AI use cases pay back fastest, and does literacy affect that?
Gartner research, as reported by CPA Practice Advisor on 24 September 2026, found that data extraction, AP and AR automation, and report creation generally return value within nine to ten months, while data management, insight generation and forecasting take longer. Literacy determines whether even the faster use cases pay back: unreviewed output that reaches a ledger or a board pack creates either rework or financial risk, both of which erode the return.
How does the reviewer's responsibility differ across group finance roles?
The group controller's primary review task is checking that eliminations, scope exceptions and top-side adjustments are correctly reflected - the model has no visibility of these unless they are explicitly supplied. FP&A reviewers must document what assumptions were passed to the model and where human judgement overrode its output. The CFO signing board-ready output should ask the preparer two questions: what did you give the model, and what did you change?
What is the most effective way to build AI review skills in a finance team?
Running review drills against closed periods with known correct answers is the fastest way to build structured scepticism without live-close pressure. Pairing senior reviewers with junior preparers for the first several periods of any new AI-assisted process transfers group-specific context that rarely survives documentation alone. Making it a team norm to record what the model was and was not given - data source, date range, prompt assumptions - creates a working paper that makes subsequent review faster and more consistent.
How can a group CFO tell that AI literacy in the team is actually improving?
The leading indicator is the quality of challenge before output is accepted: a literate team asks specific, answerable questions about model inputs rather than querying the output after the fact. The lagging indicator is the rate of material correction after first review, which should fall over three to six periods as reviewer capability improves. A flat or rising correction rate signals a literacy gap, not a tooling limitation.